IceCube Upgrade Camera System

• IceCube Upgrade Camera System is a network of integrated CMOS cameras and LED illuminators in DOMs designed for meter-scale resolution of ice properties.
• The system employs synchronized measurement campaigns and advanced image analysis, including Bayesian triangulation and machine learning, to map hole-ice and bulk-ice characteristics.
• Its precise calibration reduces systematic uncertainties in neutrino event reconstruction, leading to improvements in angular and energy resolution for IceCube detectors.

The IceCube Upgrade Camera System is a distributed, in situ optical calibration network embedded in the calibration and science Digital Optical Modules (DOMs) of the IceCube Upgrade at the South Pole. Each optical module hosts multiple high-dynamic-range CMOS cameras and narrow-band LED illuminators, designed to map the scattering, absorption, and geometric arrangements of both refrozen "hole ice" in drill columns and undisturbed Antarctic glacial ice. The system is a response to the need for meter-scale resolution in the ice model, with the goal of reducing the dominant systematic uncertainties in neutrino event reconstruction.

1. System Architecture and Hardware Design

The IceCube Upgrade Camera System is fully integrated into the IceCube Upgrade DOMs, specifically the D-Egg (dual-PMT) and mDOM (multi-PMT) designs. Each D-Egg module carries three camera-LED pairs positioned on a horizontal ring, while each mDOM houses two cameras at ±45° in the upper hemisphere, one downward-facing camera, and four LEDs positioned for upward and angular illumination (Rott et al., 24 Jul 2025, Rodan et al., 2023). All modules operate between 2160–2430 m depth, with modules vertically spaced by 3 m (D-Egg, mDOM) and string separations of 21–43 m.

Camera system components:

• Sensor: Sony IMX225LQR-C CMOS, 1312 × 993 pixels, 12 bits per pixel, peak blue-green quantum efficiency; wide-angle (≈80–120° FOV in ice) (Kang et al., 2021, Rodan et al., 2023).
• Illumination module: OSRAM Oslon SSL 80 GB CS8PM1.13, λ_peak ≈ 465–470 nm, FWHM 80°, 1–1.2 W electrical, ≈43 lm luminous output (Tönnis et al., 2021, Rott et al., 24 Jul 2025).
• Electronics: On-board digitizer and buffer, CPLD/FPGA for timing and readout, interface via SPI to DOM mainboard (Kang et al., 2019).
• Mechanical integration: Precision brackets and snap-in arrangements maintain strict alignment (<5°); all optics reside within pressure-rated glass spheres for operation at –40 °C and pressures up to 70 MPa (Kang et al., 2021).

Each camera-LED node is rigidly referenced within the pressure vessel to enable precise, model-based correction of optical axes for calibration and geometry inference (Dujmović et al., 2019, Kang et al., 2021).

2. Operational Modes, Measurement Protocols, and Data Flow

The operational workflow is organized in measurement campaigns corresponding to three principal calibration goals: (1) hole-ice mapping, (2) geometry calibration, and (3) bulk-ice property measurement (Rodan et al., 2023).

2.1 Hole-Ice Measurements

Immediately post-deployment and during freeze-in, the downward (mDOM) or paired (D-Egg) cameras capture backscatter and transmission images from local and adjacent LEDs. This yields spatially resolved maps of bubble column diameter (typically 5–10 cm), radial bubble density, and relative attenuation (Rodan et al., 2023, Kang et al., 2021).

2.2 Geometry Calibration

A run with all cameras imaging known, flashed LEDs on adjacent modules determines the module positions, LED/camera orientations, and local module tilts to sub-decimeter precision. Triangulation exploits multi-view centroiding with pinhole or corrected lens models (Rott et al., 24 Jul 2025, Rodan et al., 2023).

2.3 Bulk-Ice Calibration

Long-exposure, inter-string imaging sequences measure the transmission of light over 20–45 m baselines through undisturbed bulk ice. From the intensity and radial profile, the angular distribution, and the measured geometry, the effective scattering and absorption lengths are extracted (Rodan et al., 2023, Kang et al., 2021).

Data flow and constraints: Each DOM layer is controlled via synchronized scripts, with one camera+LED pair powered at a time to respect 1.5 Mbaud data link and a 2.3 W (camera) + 1.2 W (LED) power envelope. Raw Bayer images (2.7 MB) are transferred at ≈14 s/image, with bundling/compression reducing the rate near 5 s/image. Full measurement cycles for a string layer require 60 minutes (Rodan et al., 2023).

3. Simulation and Image Analysis Methodologies

Central to the scientific output is a suite of simulation and image-analysis tools based on the Photon Propagation Code (PPC) and the CamSim framework developed for the Upgrade (Tönnis et al., 2023, Rott et al., 24 Jul 2025).

3.1 Image Generation

Photon propagation is simulated via PPC (C/CUDA ray-tracing) with configurable absorption and scattering (λₐ, λₛ), anisotropy, layered media, and explicit geometry for camera and LED positions. Hit lists of photons on camera surfaces are post-processed into synthetic 2D images, with sensor response models, electronic noise, and optical distortions applied (Tönnis et al., 2023).

3.2 Model Fitting and Parameter Extraction

Measured and simulated images are compared using pixel-wise χ² or likelihood frameworks. Optical properties (e.g., λₛ, λₐ) are fitted via minimization: $$χ^2(λ_s,λ_a) = \sum_{i,j} \frac{(D_{i,j} − S_{i,j}(λ_s,λ_a))^2}{σ_{i,j}^2}$$ where D = data, S = simulation, σ from noise (Tönnis et al., 2023, Rott et al., 24 Jul 2025). Geometry inference uses Bayesian or maximum-likelihood approaches, leveraging known module separations and observed LED spot centroids (Rott et al., 24 Jul 2025, Kang et al., 2019).

3.3 Machine Learning-Based Analysis

Recent efforts utilize deep learning, specifically convolution–transformer hybrids (“OPTICUS”), to regress optical parameters from images. These methods demonstrate sub-percent level accuracy in scattering-length determination, outperforming classical grid-scan approaches (Rott et al., 24 Jul 2025).

4. Calibration, Validation, and Uncertainty Quantification

Validation combines laboratory acceptance testing, in situ reference measurements, and systematic error budget evaluation (Kang et al., 2021, Tönnis et al., 2021). All ≈2,200 cameras underwent 48-hour thermal cycling, dark-noise, linearity and gain, and lens-sensor alignment checks. Acceptance criteria include sub-0.5mm angular misalignment, linearity R² > 0.99, and dark noise within ±2σ (Kang et al., 2021).

In situ, calibration includes:

• Dark-frame subtraction and flat-field correction with factory-derived maps,
• Centroiding and radial profile extraction from LED images,
• Comparison to simulated image banks,
• Extraction of bubble column radius via edge-detection.

Combined statistical and systematic uncertainties in extracted bulk ice optical properties are ≲20% for λₛ, ≲12% for λₐ, with geometric positions determined to ±4 cm (horizontal) and ±15 cm (depth) (Rodan et al., 2023, Rott et al., 24 Jul 2025).

5. Mathematical Framework and Optical Modeling

IceCube Upgrade Camera data analysis relies on the Beer-Lambert law for exponential attenuation: $I(d) = I_0 \exp(-d / λ_{\text{eff}})$ with

$λ_{\text{eff}} = (1/λ_a + 1/λ_s)^{-1}$

and radial/azimuthal distributions modeled via the Henyey-Greenstein phase function: $$P(θ) = \frac{1 - g^2}{4\pi (1 + g^2 - 2g\cos θ)^{3/2}}$$ where g ≈ 0.8 for deep ice (Rodan et al., 2023, Kang et al., 2021). Triangulation uses conventional pinhole projection,

$(u, v) = f \cdot (X/Z, Y/Z)$

and least-squares fits over multiple baselines to refine the 3D DOM layout (Kang et al., 2021).

AI approaches minimize mean squared error (MSE) between predicted and simulated (true) optical properties across large training datasets, with 68% containment of relative errors typically ≲0.5% for bulk ice and ≲0.2% for hole ice (Rott et al., 24 Jul 2025).

6. Performance, Results, and Sensitivities

The system demonstrates:

• Sub-decimeter (typical ≈ 0.2–0.9 m, best ≈ 0.2 m) DOM geometry reconstruction via Bayesian triangulation (Rott et al., 24 Jul 2025, Rodan et al., 2023).
• Bulk scattering length measurements (20–100 m true range) with sub-percent bias and 68% containment <0.5% using OPTICUS (Rott et al., 24 Jul 2025).
• Hole-ice column scattering length extracted to 0.18% (68% containment), and bubble-column radii resolved to ±2 cm via edge detection (Rodan et al., 2023).
• System-level robustness with allonboard electronics verified over >10 h at 1.7 km depth, and module mechanical/electronic stability proven across the production ensemble (Kang et al., 2021, Tönnis et al., 2021).

Dominant sources of systematic uncertainty include incomplete LED beam-profile knowledge, hole-ice boundary artifacts, and omitted full optical path (vessel and gel) in simulations (Rodan et al., 2023, Rott et al., 24 Jul 2025).

7. Impact and Future Directions

The IceCube Upgrade Camera System enables direct, spatially resolved calibration of both hole and bulk glacial ice. Expected impacts include a factor-of-two reduction in systematic uncertainties on absorption and scattering coefficients, a ≈10% improvement in angular resolution, and a ≈15% gain in energy resolution for neutrino reconstruction (Kang et al., 2021). Planned developments include:

• Expanding the simulation chain to model absorption, phase function, and anisotropy jointly;
• Incorporating full optical path (glass, gel, mount);
• Domain-adapted AI models for real (versus simulated) images;
• Additional camera wavelengths for spectral ice characterization;
• Tighter timing synchronization with the main IceCube clock for absolute path-length studies (Tönnis et al., 2021, Tönnis et al., 2023, Rott et al., 24 Jul 2025).

The combined hardware–simulation–analysis paradigm constitutes the foundation for future IceCube deployments, with scalable architectures applicable to IceCube-Gen2 and other large-scale Cherenkov detectors (Tönnis et al., 2023, Kang et al., 2021).